Graft

Why we included this project

Teams that lean on Claude Code, Cursor, Codex, or Gemini in a large or unfamiliar codebase often watch their token bill climb, because the agent re-explores the repo from scratch on every task. Graft addresses that by building a graph of linked markdown files, one node per system or concept, written in plain English the way a senior engineer would explain it, and feeding that context to your coding agent so it can skip the grep-and-open loop. The structural graph is built deterministically with tree-sitter and never calls a model; the summaries are generated under your own API key with whatever provider you choose. The project publishes controlled benchmarks and a SWE-bench Verified run showing fewer tool calls, lower token spend, and equal or better correctness, so it is worth a trial if you want to cut agent costs without giving up accuracy.